Use of Pyrolysis Mass Spectrometry with Superv ised Learn ing for the Assessment of the Adulteration of Milk of Different Species

Binary m ixtures of 0± 20% cows’ m ilk with ewes’ m ilk, 0± 20% cows’ m ilk with goats’ m ilk, and 0± 5% cow s’ m ilk with goats’ m ilk were subjected to pyrolysis mass spectrometry (PyMS). For an alysis of the pyrolysis mass sp ectra so as to determ ine the p ercentage adulteration of either caprine or ov ine m ilk w ith bov ine m ilk, par tial least-squares regression (PLS), p rincipal com ponents regression (PCR) and fully interconnected feed-forward ar ti® cial neural networks (ANNs) were studied. In the latter case, the w eights were mod i® ed by u sing the standard back-prop agation algor ithm , and the nodes u sed a sigm oidal squash ing function . It was fou nd that each of the methods could be u sed to prov id e calibration models which gav e excellent pred ictions for the p ercentage adulteration with cows’ m ilk to , 1% for samples, w ith an accuracy of 6 0.5% , on which they had not been trained. Scaling the in dividual nodes on the inpu t layer of ANNs sign i® cantly d ecreased the time taken for the ANNs to learn, com pared with scaling across the whole mass range; however in one case this approach resulted in poor generalization for th e estimates of percentage cows’ m ilk in ewes’ m ilk. To assess whether the calibration models had learned th e differences between the m ilk species or th e d ifferences due to the different fat conten t of in each of the m ilk types, we also analyzed pu re m ilk samples varyin g in fat conten t by PyMS. Cluster an alysis showed unequiv ocally that th e m ajor var iation between the d ifferent m ilk species was not due to var iab le fat conten t. Sin ce an y biological mater ial can be pyrolyzed in this way, the combination of PyMS with superv ised learning constitu tes a rap id , powerful, an d novel approach to the quantitativ e assessm ent of food ad ulteration generally.